AI SDK 7.0.94: ai removes the seven experimental text batch exports

ai@7.0.94 is a patch release that reshapes the experimental batch API. Seven names leave the package and nine arrive in their place: a batch is no longer text-specific, and the model it runs against is now named per request rather than once for the whole batch. The package’s changelog files the work under “feat: support per-request models in batch.” Only code that imports one of the seven removed names from ai is affected.

What changed

• experimental_startTextBatch is now experimental_startBatch, and it takes a provider rather than a model. Experimental_StartTextBatchOptions and Experimental_StartTextBatchResult go with it, replaced by Experimental_StartBatchOptions and Experimental_StartBatchResult. The new options carry an optional provider, which defaults to the global provider or to the Vercel AI Gateway, in place of the required model.
• The batch-level model, tools, toolChoice, toolOrder and toolsContext move onto each request. A request is now discriminated by modality: Experimental_TextBatchRequest keeps its name but requires type: 'text' and its own model, and Experimental_BatchRequest is the union over modalities.
• Experimental_TextBatch and Experimental_TextBatchReference are gone. Experimental_Batch and Experimental_BatchReference replace them. The reference is a new shape as well as a new name: version is 2, and the type and modelId fields are no longer part of it, since a batch can hold requests against more than one model.
• Experimental_BatchOperationOptions splits in two. experimental_getBatchStatus and experimental_getBatchResults keep their names, but now take Experimental_GetBatchStatusOptions and Experimental_GetBatchResultsOptions, each naming a provider instead of a model.
• Experimental_BatchLanguageModel is gone. Nothing replaces the type directly. A model is a plain model id string on a request, and the provider is typed by Experimental_BatchProvider.

Before and after

The import Emendant’s matchers look for, one per removed name:

import { experimental_startTextBatch } from 'ai';

export { experimental_startTextBatch };

No transform can write any of the seven, so there is no rewritten file to show. An assisted patch is offered for each, and what it has to work out follows from the two option types. In ai@7.0.93, the model and the tools belonged to the batch:

type StartTextBatchOptions<TOOLS extends ToolSet = ToolSet> = {
    model: BatchLanguageModel;
    requests: ReadonlyArray<TextBatchRequest>;
    tools?: TOOLS;
    toolChoice?: ToolChoice<NoInfer<TOOLS>>;
    providerOptions?: ProviderOptions;
    webhookUrl?: string;
} & BatchRequestOptions;

In ai@7.0.94, the batch names a provider and each request names its own model:

type StartBatchOptions<TOOLS extends ToolSet = ToolSet, PROVIDER extends BatchProvider = BatchProvider> = {
    provider?: PROVIDER;
    requests: ReadonlyArray<BatchRequest<InferBatchModelIds<PROVIDER>, TOOLS>>;
    providerOptions?: ProviderOptions;
    webhookUrl?: string;
} & BatchCallOptions;

Migration guidance and its limits

The seven removed names carry an assisted fix. Name a model provider on the command line or in emendant.json and emendant fix tries each one; without one they are reported with their guidance and nothing is sent anywhere. Expect declines among them: the guidance names the construct that went and tells the model to set fixed to false rather than invent a replacement, and several of these have no replacement that is a rename. Experimental_BatchLanguageModel has no direct successor at all, and experimental_startTextBatch becomes a differently shaped call rather than a differently spelled one. A patch that is written is proved in a copy of your repository moved to 7.0.94 and put through whichever of your own typecheck and tests it finds there, and withheld if it fails.

The two findings on experimental_getBatchStatus and experimental_getBatchResults are report only. Both fire at the options literal rather than at an import, because the model key those calls took is the thing that went, and what belongs there instead is a provider on a differently named type rather than a key to rename.

Rename the call and the two option types, then move what the batch used to hold down onto the requests. Each entry in requests takes type: 'text' and a model, plus the tools, toolChoice, toolOrder and toolsContext that used to sit beside requests. Drop Experimental_BatchLanguageModel from your own signatures and use the model id string the request expects.

Persisted references do not survive the upgrade. A batch reference stored by ai@7.0.93 is version 1 and carries type and modelId; the code reading it in 7.0.94 expects version 2 and neither field. A batch started before the upgrade has to be drained before it, or its reference migrated by hand.

The matchers fire on the import statement alone. Two things a scan therefore stays quiet about: Experimental_BatchReference and Experimental_TextBatchRequest kept their names while changing shape, so importing either is not a finding even though the code around it may no longer compile. A clean scan says a file does not import these seven names from ai, and nothing more about the release.

What Emendant detects, fixes and verifies

9 changes in ai 7.0.94. 0 patched by a transform, 7 patched only with a model provider you name, 2 report only. Every patch is proved in a copy of your repository before it is offered, and its header names the checks that passed. This table is generated from the feed entry, so it cannot claim more than the entry does.

emendant explain <change-id> prints any row's entry, guidance and sources at the terminal.

Coverage

Snapshot

2026-09-20.1, sequence 27, signed 20 September 2026

Minimum CLI

emendant@0.1.4

Feed entry

feed/npm/ai/7.0.94.json

emendant feed status shows the snapshot your machine holds, and emendant feed update fetches the latest.

Primary sources

Every claim above was checked against these, each pinned to the release rather than to a default branch.

• ai@7.0.94npm registry, the published package
• ai@7.0.93npm registry, the published package

Scan your repository

Runs locally, reads your lockfile and source, sends nothing anywhere.

npx emendant scan

Then npx emendant fix writes the patches the table says exist. The getting started guide covers the flags and the patch grades.